70% of the Web: AI Models Expose Critical PHP Flaws That Eluded Humans for Two Years

OpenAI announced on October 21 that its GPT-5.6 Sol model played a central role in uncovering critical security vulnerabilities in PHP, the open-source programming language that powers an estimated 70% of the internet. The discovery, made under the company’s “Patch the Planet” cybersecurity initiative, has triggered a coordinated response involving PHP maintainers, security providers, and government agencies to deploy fixes before the weaknesses can be widely exploited by malicious actors.

The situation escalated rapidly due to the ubiquity of the affected software. While the underlying code defects had been known to the PHP Foundation since 2024, human developers were unable to determine a reliable method for exploitation or a definitive fix. This stalemate persisted for two years until AI-assisted researchers from Calif, a California-based cybersecurity firm and founding partner of the initiative, identified an exploitable version of the bug on August 11. OpenAI’s own researchers confirmed the vulnerability the same day, while another firm, Wiz, independently discovered a second related flaw shortly thereafter.

The two vulnerabilities, designated as CVE-2026-109618 and CVE-2026-109619, both score a critical severity rating of 9.2 out of 10. They stem from flaws in how the PHP engine manages memory and processes errors. Specifically, the first issue occurs when the system runs out of allocated memory, potentially leaving incomplete data that the engine later misinterprets as valid. The second involves a fatal error that causes the engine to point to memory that has already been freed, allowing subsequent requests to access or corrupt that space.

Because these flaws reside in the core engine of PHP, they affect websites even if the specific application code is secure. Attackers can exploit these weaknesses remotely without requiring login credentials, potentially allowing them to execute malicious code and take control of affected servers. Furthermore, the AI-assisted analysis demonstrated that the bugs could expose sensitive data stored in server memory, including passwords, authentication tokens, and API keys.

The PHP Foundation released patched versions 8.2.35, 8.3.36, 8.4.27, and 8.5.12 to address the issues. Versions 8.1 and older will not receive official fixes, meaning users of those iterations must upgrade to remain secure. The coordinated release was originally scheduled for October 22 but was moved up to October 21 after premature disclosures on October 20 led to public concern. OpenAI states that the two-month timeline for the fix was necessary due to the complexity of the task, which required significant changes to the PHP engine.

OpenAI is framing this incident as a prime example of its “Defender's Window” concept. Introduced after the company’s models hacked Hugging Face, this theory posits a narrow timeframe during which organizations can use advanced AI to find and patch defenses before those same capabilities become widely available to attackers. The company argues that since powerful technology will inevitably be used by bad actors, it is a collective responsibility to act proactively. This view contrasts with the argument that these flaws were effectively harmless because they were not exploitable without AI assistance.

The initiative began on July 30, 2026, when Calif researchers started auditing PHP using OpenAI tools at the request of the PHP Foundation. Within two weeks, the AI-wielding team demonstrated practical exploitation methods for the two-year-old bug that had previously stumped human security teams. OpenAI considers this one of the most significant cybersecurity discoveries of 2026, highlighting the shifting dynamic between offensive and defensive AI capabilities.

Other industry players are seeing similar trends. Anthropic’s Claude Mythos recently identified thousands of zero-day vulnerabilities in major operating systems and web browsers. As these AI-driven discoveries become more frequent, the pressure on legacy software foundations to upgrade their testing and patching processes intensifies. For website administrators, the immediate takeaway is clear: upgrading to the latest patched PHP versions is now essential to mitigate the risk of remote code execution and data exposure.

The incident underscores the growing reliance on AI to audit complex codebases that have evolved over decades. While the speed at which the exploits were found raises questions about the longevity of current security practices, it also demonstrates the potential for AI to accelerate the defensive cycle. The next few months will likely see increased scrutiny of other widely used open-source projects as researchers apply similar AI-assisted auditing techniques.